Month: December 2016

This is a second technical post related to segment-routing, I did a basic introduction to this technology on Juniper MX here;

https://tgregory.org/2016/08/13/segment-routing-on-junos-the-basics/

For this post I’m looking at something a bit more advanced and fun – performing Segment-routing traffic-engineering using an SDN controller, in this case OpenDaylight Beryllium – an open source SDN controller with some very powerful functionality.

This post will use Cisco ASR9kV virtual routers running on a Cisco UCS chassis, mostly because Cisco currently have the leading-edge support for Segment-routing at this time, Juniper seem to be lagging behind a bit on that front!

Lets check out the topology;

It’s a pretty simple scenario – all of the routers in the topology are configured in the following way;

• XRV-1 to XRV-8; PE routers (BGP IPv4)
• XRV 2 to XRV7; P routers (ISIS-Segment-routing)
• XRV4 is an in-path RR connecting to the ODL controller

The first thing to look at here is BGP-LS “BGP Link-state” which is an extension of BGP that allows IGP information (OSPF/ISIS) to be injected into BGP, this falls conveniently into the world of centralised path computation – where we can use a controller of some sort to look at the network’s link-state information, then compute a path through the network. The controller can then communicate that path back down to a device within the network using a different method, ultimately resulting in an action of some sort – for example, signalling an LSP.

Some older historic platforms such as HP Route analytics – which enabled you to discover the live IGP topology by running ISIS or OSPF directly with a network device, however IGPs tend to be very intense protocols and also require additional effort to support within an application, rather than a traditional router. IGPs are only usually limited to the domain within which they operate – for example if we have a large network with many different IGP domains or inter-domain MPLS, the IGP’s view becomes much more limited. BGP on the other hand can bridge many of these gaps, and when programmed with the ability to carry IGP information – can be quite useful.

The next element is PCE or Path computation element – which generally contains two core elements;

• PCC – Path computation client – In the case of this lab network, a PCC would be a PE router
• PCE – Path computation element – In the case of this lab network, the PCE would be the ODL controller

These elements communicate using PCEP (Path computation element protocol) which allows a central controller (in this case ODL) to essentially program the PCC with a path – for example, by signalling the actual LSP;

Basic components;

Basic components plus an application (in this case Pathman-SR) which can compute and signal an LSP from ODL to the PCC (XRV-1);

In the above example, an opensource application (in this case Pathman-SR) is using the information about the network topology obtained via BGP-LS and PCE, stored inside ODL – to compute and signal a Segment-routing LSP from XRV-1 to XRV-8, via XRV3, XRV5 and XRV7.

Before we look at the routers, lets take a quick look at OpenDaylight, general information can be found here; https://www.opendaylight.org I’m running Beryllium 0.4.3 which is the same Cisco’s DCloud demo – it’s a relatively straightforward install process, I’m running my copy on top of a standard Ubuntu install.

From inside ODL you can use the YANG UI to query information held inside the controller, which is essentially a much easier way of querying the data, using presets – for example, I can view the link-state topology learnt via BGP-LS pretty easily;

There’s a whole load of functionality possible with ODL, from BGP-Flowspec, to Openflow, to LSP provisioning, for now we’re just going to keep it basic – all of this is opensource and requires quite a bit of “playing” to get working.

Lets take a look at provisioning some segment-routing TE tunnels, first a reminder of the diagram;

And an example of some configuration – XRv-1

ISIS;

A relatively simple ISIS configuration, with nothing remarkable going on,

• Line 9 enabled Segment-Routing for ISIS
• Line 13 injects a SID (Segment-identifier) of 10 into ISIS for loopback 0

The other aspect of the configuration which generates a bit of interest, is the PCE and mpls traffic-eng configuration;

• Line 1 enables basic traffic-engineering, an important point to note – to do MPLS-TE for Segment-routing, you don’t need to turn on TE on every single interface like you would if you were using RSVP, so long as ISIS TE is enabled and
• Lines 2, 3 and 4 connect the router from it’s loopback address, to the opendaylight controller and enable PCE
• Line 6 through 9 specify the segment-routing parameters for TE
• Line 14 specifies the tunnel ID for automatically generated tunnels – for tunnels spawned by the controller

Going back to the diagram, XRv-4 was also configured for BGP-LS;

• Line 6 enables the BGP Link-state AFI/SAFI
• Lines 8 through 19 are standard BGP RR config for IPv4
• Line 22 is the BGP peer for the Opendaylight controller
• Line 28 turns on the link-state AFI/SAFI for Opendaylight

Also of Interest on XRv-4 is the ISIS configuration;

• Line 4 copies the ISIS link-state information into BGP-link state

If we do a “show bgp link-state link-state” we can see the information taken from ISIS, injected into BGP – and subsequently advertised to Opendaylight;

With this information we can use an additional app on top of OpenDaylight to provision some Segment-routing LSPs, in this case I’m going to use something from Cisco Devnet called Pathman-SR – it essentially connects to ODL using REST to program the network, Pathman can be found here; https://github.com/CiscoDevNet/pathman-sr

Once it’s installed and running, simply browse to it’s url (http://192.168.3.250:8020/cisco-ctao/apps/pathman_sr/index.html) and you’re presented with a nice view of the network;

From here, it’s possible to compute a path from one point to another – then signal that LSP on the network using PCEP, in this case – lets program a path from XRv9k-1 to XRv9k-8

In this case, lets program a path via XRV9k-2, via 4, via 7 to 8;

Once Pathman has calculated the path – hit deploy, Pathman sends the path to ODL – which then connects via PCEP to XRV9kv-1 and provisions the LSP;

Once this is done, it’s check XRV9k-1 to check out the SR-TE tunnel;

We can see from the output of “show ip int brief” on line 5, that interface tunnel-te1 has been created, but it’s nowhere in the config;

PCE signalled LSPs never appear in the configuration, they’re created, managed and deleted by the controller – it is possible to manually add an LSP then delegate it to the controller, but that’s beyond the scope here (that’s technical speak for “I couldn’t make it work 🙂 )

Lets check out the details of the SR-TE tunnel;

Points of interest;

• Line 4 shows the name of the LSP as configured by Pathman
• Line 7 shows that the signalling is Segment-routing via autoPCC
• Lines 33 and 34 show the tunnel was generated by the Opendaylight controller
• Lines 39 shows the LSP is PCE controlled
• Lines 40 through 43 show the programmed path
• Line 44 basically shows XRV9k-1 being the SR-TE headend,

Lines 40-43 show some of the main benefits of Segment-routing, we have a programmed traffic-engineered path through the network, but with far less control-plane overhead than if we’d done this with RSVP-TE, for example – lets look at the routers in the path (xrv-2 xrv-4 and xrv-7)

Essentially – the path that the SR-TE tunnel takes contains no real control-plane state, this is a real advantage for large networks as the whole thing is much more efficient.

The only pitfall here, is that whilst we’ve generated a Segment-routed LSP, like all MPLS-TE tunnels we need to tell the router to put traffic into it – normally we do this with autoroute-announce or a static route, at this time OpenDaylight doesn’t support the PCEP extensions to actually configure a static route, so we still need to manually put traffic into the tunnel – this is fixed in Cisco’s openSDN and WAE (wan automation engine)

I regularly do testing and development work with some of the largest ISPs in the UK – and something that regularly comes up, is where customers are running a traditional full-mesh of RSVP LSPs, if you have 500 edge routers – that’s 250k LSPs being signalled end to end, the “P” routers in the network need to signal and maintain all of that state. When I do testing in these sorts of environments, it’s not uncommon to see nasty problems with route-engine CPUs when links fail, as those 250k LSPs end up having to be re-signalled – indeed this very subject came up in a conversation at LINX95 last week.

With Segment-routing, the traffic-engineered path is basically encoded into the packet with MPLS labels – the only real difficulty is that it requires the use of more labels in the packet, but once the hardware can deal with the label-depth, I think it’s a much better solution than RSVP, it’s more efficient and it’s far simpler.

From my perspective – all I’ve really shown here is a basic LSP provisioning tool, but it’s nice to be able to get the basics working, in the future I hope to get my hands on a segment-routing enabled version of Northstar, or Cisco’s OpenSDN controller – (which is Cisco productised version of ODL) 🙂